Chapter 19
Models of
Transactions
Structuring Applications
• Many applications involve long transactions
that make many database accesses
• To deal with such complex applications
many transaction processing systems
provide mechanisms for imposing some
structure on transactions
2
Flat Transaction
• Consists of:
– Computation on local variables
begin transaction
• not seen by DBMS; hence will be
ignored in most future discussion
– Access to DBMS using call or
statement level interface
• This is transaction schedule; commit
applies to these operations
• No internal structure
• Accesses a single DBMS
• Adequate for simple applications
EXEC SQL …..
EXEC SQL …..
commit
3
Flat Transaction
• Abort causes the
execution of a program
that restores the
variables updated by the
transaction to the state
they had when the
transaction first accessed
them.
begin transaction
EXEC SQL …..
EXEC SQL …..
if condition then abort
commit
4
Some Limitations of Flat
Transactions
• Only total rollback (abort) is possible
– Partial rollback not possible
• All work lost in case of crash
• Limited to accessing a single DBMS
• Entire transaction takes place at a single
point in time
5
Providing Structure Within a
Single Transaction
6
Savepoints
• Problem: Transaction detects condition that
requires rollback of recent database changes
that it has made
• Solution 1: Transaction reverses changes
itself
• Solution 2: Transaction uses the rollback
facility within DBMS to undo the changes
7
Savepoints
begin transaction
S1;
Call to DBMS
sp1 := create_savepoint();
S2;
sp2 := create_savepoint();
S3;
if (condition) {rollback (sp1); S5};
S4;
commit
• Rollback to spi causes database updates subsequent to creation
of spi to be undone
– S2 and S3 updated the database (else there is no point
rolling back over them)
• Program counter and local variables are not rolled back
• Savepoint creation does not make prior database changes
8
durable (abort rolls all changes back)
Example of Savepoints
• Suppose we are making airplane
reservations for a long trip
– London-NY NY-Chicago Chicago-Des Moines
• We might put savepoints after the code that made
the London-NY and NY-Chicago reservations
• If we cannot get a reservation from Chicago to
Des Moines, we would rollback to the savepoint
after the London-NY reservation and then perhaps
try to get a reservation through St Louis
9
Distributed Systems: Integration
of Legacy Applications
• Problem: Many enterprises support multiple
legacy systems doing separate tasks
– Increasing automation requires that these systems be
integrated
withdraw part
return part
stock level
Inventory
Application
DBMS 1
Site B
order part
payment
Billing
Application
DBMS 2
Site C
10
Distributed Transactions
• Incorporate transactions at multiple servers into a
single (distributed) transaction
– Not all distributed applications are legacy systems;
some are built from scratch as distributed systems
tx_begin;
order_part;
withdraw_part;
payment;
tx_commit;
Inventory
Application
DBMS 1 Site B
Billing
Application
DBMS 2 Site C
Site A
11
Distributed Transactions
• Goal: distributed transaction should be ACID
– Each subtransaction is locally ACID (e.g., local
constraints maintained, locally serializable)
– In addition the transaction should be globally ACID
• A: Either all subtransactions commit or all abort
• C: Global integrity constraints are maintained
• I: Concurrently executing distributed transactions are
globally serializable
• D: Each subtransaction is durable
12
Banking Example
• Global atomicity - funds transfer
– Either both subtransactions commit or neither does
tx_begin;
withdraw(acct1);
deposit(acct2);
tx_commit;
13
Banking Example (con’t)
• Global consistency – Sum of all account balances at bank branches =
total assets recorded at main office
14
Banking Example (con’t)
• Global isolation - local serializability at each site
does not guarantee global serializability
– post_interest subtransaction is serialized after audit
subtransaction in DBMS at branch 1 and before audit
in DBMS at branch 2 (local isolation), but
– there is no global order
post_interest
time

audit
sum balances at branch 1;
post interest at branch 1;
post interest at branch 2;
sum balances at branch 2;
15
Exported Interfaces
Local system might export an interface for executing
individual SQL statements.
subtransaction
tx_begin;
EXEC SQL SELECT..
EXEC SQL INSERT..
EXEC SQL SELECT..
tx_commit;
site A
DBMS 1
site B
DBMS 2
site C
subtransaction
Alternatively, the local system might export an interface
for executing subtransactions.
16
Multidatabase
• Set of databases accessed by a distributed
transaction is referred to as a multidatabase (or
federated database)
– Each database retains its autonomy and might support
local (non-distributed) transactions
• Multidatabase might have global integrity
constraints
– e.g., Sum of balances of individual bank accounts at all
branch offices = total assets stored at main office
17
Transaction Hierarchy
• A distributed transaction invokes subtransactions.
• General model: one distributed transaction might
invoke another as a subtransaction, yielding a
hierarchical structure
Distributed transactions
18
Models of Distributed Transactions
• Can siblings execute
concurrently?
• Can parent execute concurrently
with children?
• Who initiates commit?
Hierarchical Model: No concurrency among subtransactions, root
initiates commit
Peer Model: Concurrency among siblings and between parent and
children, any subtransaction can initiate commit
19
Distributed Transactions
• Transaction designer has little control over
the structure. Decomposition fixed by
distribution of data and/or exported
interfaces (legacy environment)
• Essentially a bottom-up design
20
Nested Transactions
• Problem: Lack of mechanisms that allow:
– a top-down, functional decomposition of a
transaction into subtransactions
– individual subtransactions to abort without
aborting the entire transaction
• Although a nested transaction looks similar
to a distributed transaction, it is not
conceived of as a tool for accessing a
multidatabase
21
Characteristics of Nested Transactions
• (1) Parent can create children to
perform subtasks; children might
execute sequentially or concurrently;
parent waits until all children
complete (no communication
between parent and children).
• (2) Each subtransaction (together with its descendants) is
isolated with respect to each sibling (and its descendants).
Hence, siblings are serializable, but order is not determined
and nested transaction is non-deterministic.
• (3) Concurrent nested transactions are serializable.
22
Characteristics of Nested Transactions
• (4) A subtransaction is atomic. It can
abort or commit independently of other
subtransactions. Commit is conditional
on commit of parent (since child task is a
subtask of parent task). Abort causes
abort of all subtransaction’s children.
• (5) Nested transaction commits when root commits. At
that point updates of committed subtransactions are made
durable.
23
Nested Transaction - Example
Booking a flight from
London to Des Moines
L -- DM
C
concurrent
C L -- NY
C
C = commit
A = abort
NY -- DM
sequential
NY -- Chic -- DM A
stop in Chicago
C/A
NY -- Chic
concurrent
concurrent
A
Chic -- DM
C NY -- StL -- DM
C
NY -- StL
stop in St. Louis
C
StL -- DM 24
Nested Transactions
parent of all
nested transactions
concurrent
isolation
isolation
isolation
25
Characteristics of Nested
Transactions
• (6) Individual subtransactions are not necessarily
consistent, but nested transaction as a whole is
consistent
26
Structuring to Increase
Transaction Performance
• Problem: In the models previously discussed, a
transaction generally locks items it accesses and holds
locks until commit time to guarantee serializabiltiy
acquire lock on x
release lock on x


T1: r(x:12) .. compute .. w(x:13) commit
T2:
request read(x)

(wait)
r(x:13) ..compute.. w(x:14) ..

acquire lock on x
– This eliminates bad interleavings, but limits concurrency and
hence performance
27
Example - Switch Sections
transaction moves
student from section
s1 to section s2,
uses TestInc, Dec
Move(s1, s2)
Section abstr.
L2 TestInc(s2)
Dec(s1)
Tuple abstr.
L1 Sel(t2)
Upd(t2)
enrollments
stored in tuples
t1 and t2
Upd(t1)
Page abstr.
L0
Rd(p2)
Rd(p2) Wr(p2)
time
tuples stored
in pages
p1 and p2
Rd(p1)
Wr(p1)
28
Structuring into Multiple
Transactions
29
Chained Transactions
• Problem 1 (trivial): Invoking
begin_transaction at the start of each
transaction involves communication
overhead
• With chaining, a new transaction is started
automatically for an application program
when the program commits or aborts the
previous one
– This is the approach taken in SQL
30
Chained Transactions
begin transaction
S1
commit
S2
begin transaction
S3
commit
S2 not included in a
transaction since it
has no db operations
begin transaction
S1
commit
begin transaction
S2
S3
commit
Equivalent since
S2 does not access
the database
transaction
starts
implicitly
S1
commit
S2
S3
commit
Chaining
equivalent
31
Chained Transactions
• Problem 2: If the system crashes during the
execution of a long-running transaction,
considerable work can be lost
• Chaining allows a transaction
to be decomposed into subtransactions with intermediate
commit points
• Database updates are made
durable at intermediate points
=> less work is lost in a crash
S1
=>
S2
S3
commit
S1;
commit;
S2;
commit;
S3;
commit;
32
Example
S1;
-- update recs 1 - 1000
commit;
S2;
-- update recs 1001 - 2000
commit;
S3;
-- update recs 2001 - 3000
commit;
• Chaining compared with savepoints:
– Savepoint: explicit rollback to arbitrary savepoint; all
updates lost in a crash
– Chaining: abort rolls back to last commit; only the
updates of the most recent transaction lost in a crash
33
Chaining Considerations Atomicity
• Transaction as a whole is not atomic. If
crash occurs
– DBMS cannot roll the entire transaction back
• Initial subtransactions have committed,
– Their updates are durable
– The updates might have been accessed by other
transactions (locks have been released)
– Hence, the application must roll itself forward
34
Chaining Considerations Atomicity
• Roll forward requires that on recovery the application
can determine how much work has been committed
– Each subtransaction must tell successor where it left off
• Communication between successive subtransactions
cannot use local variables (they are lost in a crash)
– Use the database to communicate between subtransactions
r(rec_index:0);
S1;
w(rec_index:1000);
commit;
r(rec_index:1000);
S2;
w(rec_index:2000);
commit;
-- update records 1 - 1000
-- save index of last record updated
-- get index of last record updated
-- update records 1001 – 2000
35
Chaining Considerations
• Transaction as a whole is not isolated.
– Database state between successive subtransactions might
change since locks are released (but performance improves)
subtransaction 1
subtransaction 2
T1: r(x:15)…w(x:24)… commit
r(x:30)…
T2:
…w(x:30)…commit
• Subtransactions might not be consistent
– Inconsistent intermediate states visible to concurrent
transactions during execution or after a crash
36
Alternative Semantics for
Chaining
S1;
chain;
S2;
chain;
S3;
commit;
• Chain commits the transaction (makes it durable) and
starts a new transaction, but does not release locks
– Individual transactions do not have to be consistent
– Recovery is complicated (as before); rollforward required
– No performance gain
37
A Problem With Obtaining
Atomicity With Chaining
• Suppose we use the first semantics for chaining
– Subtransactions give up locks when they commit
• Suppose that after a subtransaction of a transaction
T makes its changes to some item and commits
– Another transaction changes the same item and
commits
– T would then like to abort
– Based on our usual definition of chained transactions,
atomicity cannot be achieved because of the committed
subtransactions
38
Partial Atomicity
• Suppose we want to achieve some measure
of atomicity by undoing the effects of all
the committed subtransactions when the
overall transaction wants to abort
• We might think we can undo the updates
made by T by just restoring the values each
item had when T started (physical logging)
– This will not work
39
An Example
T1: Update(x)1,1 commit1,1
…
abort1
T2:
Update(x) commit
If, when T1 aborts, we just restore the value of x to
the value it had before T1 updated it, T2’s update
would be lost
40
Compensation
• One approach to this problem is compensation
• Instead of restoring a value physically, we restore
it logically by executing a compensating
transaction
– In the student registration system, a Deregistration
subtransaction compensates for a successful
Registration subtransaction
– Thus Registration increments the Enrollment attribute
and Deregistration decrements that same attribute
• Compensation works even if some other concurrent
Registration subtransaction has also incremented Enrollment
41
Sagas: An Extension To Chained
Transactions That Achieves
Partial Atomicity
• For each subtransaction, STi,j in a chained
transaction, Ti a compensating transaction, CTi is
designed
• Thus if a transaction T1 consisting of 5 chained
subtransactions aborts after the first 3
subtransactions have committed, then
ST1,1ST1,2ST1,3CT1,3CT1,2CT1,1
will perform the desired compensation
42
Sagas and Atomicity
• With this type of compensation, when a
transaction aborts, the value of every item it
changed is eventually restored to the value
it had before that transaction started
• However, complete atomicity is not
guaranteed
– Some other concurrent transaction might have
read the changed value before it was restored to
its original value
43
Declarative Transaction
Demarcation
• We have already talked about two ways in
which procedures can execute within a
transaction
– As a part of the transaction
• Stored procedure
– As a child in a nested transaction
44
Declarative Transaction
Demarcation (con’t)
• Two other possible ways
– The calling transaction is suspended, and a new
transaction is started. When it completes the first
transaction continues
• Example: The called procedure is at a site that charges for its
services and wants to be paid even if the calling transaction
aborts
– The calling transaction is suspended, and the called
procedure executes outside of any transaction. When it
completes the first transaction continues
• Example: The called procedure accesses a non-transactional
file system
45
Declarative Transaction
Demarcation (con’t)
• One way to implement such alternatives is
through declarative transaction
demarcation
– Declare in some data structure, outside of any
transaction, the desired transactional behavior
– When the procedure is called, the system
intercepts the call and provides the desired
behavior
46
Implementation of Declarative
Transaction Demarcation
• Declarative transaction demarcation is
implemented within J2EE and .NET
– We discuss J2EE (.NET is similar)
• The desired transactional behavior of each
procedure is declared as an attributed in a
separate file called the deployment
descriptor
47
Transaction Attributes
• Possible attributes (in J2EE) are
–
–
–
–
–
–
Required
RequiresNew
Mandatory
NotSupported
Supports
Never
• The behavior for each attribute depends on
whether or not the procedure is called from within
a procedure
– All possibilities are on the next slide
48
Status of Calling Method
Attribute of Called Not in a
Method
Transaction
In a
Transaction
Required
Starts a New
Transaction
Executes Within
the Transaction
RequiresNew
Starts a New
Transaction
Starts a New
Transaction
Mandatory
Exception
Thrown
Executes Within
the Transaction
NotSupported
Transaction Not
Started
Transaction
Suspended
Supports
Transaction Not
Started
Executes Within
the Transaction
Never
Transaction Not
Started
Exception
Thrown
All Possibilities
49
Description of Each Attribute
• Required:
– The procedure must execute within a
transaction
• If called from outside a transaction, a new
transaction is started
• If called from within a transaction, it executes
within that transaction
50
Description (con’t)
• RequiresNew:
– Must execute within a new transaction
• If called from outside a transaction, a new
transaction is started
• If called from within a transaction, that transaction
is suspended and a new transaction is started.
When that transaction completes, the first
transaction resumes
– Note that this semantics is different from nested
transactions. In this case the commit of the new
transaction is not conditional.
51
Description (con’t)
• Mandatory:
– Must execute within an existing transaction
• If called from outside a transaction, an exception is
thrown
• If called from within a transaction, it executes
within that transaction
52
Description (con’t)
• NotSupported:
– Does not support transaction
• If called from outside a transaction, a transaction is
not started
• If called from inside a transaction, that transaction is
suspended until the procedure completes after which
the transaction resumes
53
Description (con’t)
• Supports:
– Can execute within or not within a transaction,
but cannot start a new transaction
• If called from outside a transaction, a transaction is
not started
• If called from inside a transaction, it executes within
that transaction
54
Description (con’t)
• Never:
– Can never execute within a transaction
• If called from outside a transaction, a new
transaction is not started
• If called from within a transaction, an exception is
thrown
55
Example
• The Deposit and Withdraw transactions in
a banking application would have attribute
Required.
– If called to perform a deposit, a new transaction
would be started
– If called from within a Transfer transaction to
transfer money between accounts, they would
execute within that transaction
56
Advantages
• Designer of individual procedures does not have
to know the transactional context in which the
procedure will be used
• The same procedure can be used in different
transaction contexts
– Different attributes are specified for each different
context
• We discuss J2EE in more detail and how
declarative transaction demarcation is
implemented in J2EE in the Architecture chapter.
57
Multilevel Transactions
• A multilevel transaction is a nested set of
subtransactions.
– The commitment of a subtransaction is
unconditional, causing it to release its locks, but
– Multilevel transactions are atomic and their
concurrent execution is serializable
58
Multilevel Transactions
• Data is viewed as a sequence of increasing,
application oriented, levels of abstraction
• Each level supports a set of abstract objects
and abstract operations (methods) for
accessing those objects
• Each abstract operation is implemented as a
transaction using the abstractions at the next
lower level
59
Example - Switch Sections
Move(s1, s2)
transaction (sequential),
moves student from one
section to another,
uses TestInc, Dec
Section abstr.
L2 TestInc(s2)
Dec(s1)
Tuple abstr.
L1 Sel(t2)
Upd(t2)
Upd(t1)
Page abstr.
L0
Rd(p2)
Rd(p2) Wr(p2)
time
Rd(p1)
Wr(p1)
60
Multilevel Transactions
• Parent initiates a single subtransaction at a
time and waits for its completion. Hence a
multilevel transaction is sequential.
• All leaf subtransactions in the tree are at the
same level
• Only leaf transactions access the database.
• Compare with distributed and nested
models
61
Multilevel Transactions
• When a subtransaction (at any level) completes, it
commits unconditionally and releases locks that it
has acquired on items at the next lower level.
– TestInc(s2) locks t2; unlocks t2 when it commits
• The change it has made to the locked item becomes
visible to subtransactions of other transactions
– The incremented value of t2 is visible to a subsequent
execution of TestInc or Dec by concurrent transactions
• This creates problems maintaining isolation and
atomicity.
62
Maintaining Isolation
p2 is unlocked
when Sel commits

TestInc1: Sel(t2)
Upd(t2)
TestInc2:
Sel(t2) Upd(t2)

Sel2 can lock p2
• Problem: Interleaved execution of two
TestInc’s results in error (we will return to this
later)
63
Maintaining Atomicity
Move1: TestInc(s2) Dec(s1)
abort
Move2:
TestInc(s3) Dec(s1) commit
• When T1 aborts, the value of s1 that existed prior to its
access cannot simply be restored (physical restoration)
• Logical restoration must be done using compensating
transactions
– Inc compensates for Dec; Dec compensates for a
successful TestInc; no compensation needed for
unsuccessful TestInc
64
Compensating Transactions
• Multilevel model uses compensating
transaction
logical restoration
(using compensation)
caused by abort
T1: TestInc(s2) Dec(s1)
Inc(s1) Dec(s2)
T2:
TestInc(s3) Dec(s1) commit
65
Correctness of Multilevel
Transactions
• As we shall see later,
– Multilevel transactions are atomic
• In contrast with Sagas, which also use
compensation, but do not guarantee atomicity
– Concurrent execution of multilevel transactions
is serializable
66
Recoverable Queues
• Problem: Distributed model assumes that the
subtransactions of a transaction follow one
another immediately (or are concurrent).
• In some applications the requirement is that a
subtransaction be eventually executed, but not
necessarily immediately.
• A recoverable queue is a transactional data
structure in which information about
transactions to be executed later can be durably
stored.
67
Transactional Features
T1: begin transaction;
compute;
item := service
description
enqueue(item);
commit;
recoverable
queue
T2: begin transaction;
dequeue(item);
perform requested
service;
commit;
• Item is enqueued if T1 commits (deleted if it aborts); item
is deleted if T2 commits (restored if it aborts)
• An item enqueued by T1 cannot be dequeued by T2 until
T1 commits
• Queue is durable
68
Pipeline Queue for Billing
Application
shipping
queue
order entry
transaction
billing
queue
shipping
transaction
billing
transaction
69
Concurrent Implemention of the
Same Application
shipping
queue
shipping
transaction
billing
queue
billing
transaction
order entry
transaction
70
Recoverable Queue
• Queue could be implemented within database,
but performance suffers
– A transaction should not hold long duration locks
on a heavily used data structure
acquire lock on queue in db
release lock on queue


T1: enq(I1) …………..compute………………commit
T2:
request enq(I2) (wait)
enq(I2)
T3:
request enq(I3) (wait)
T4:
request enq(I4) (wait)
acquire lock on queue
71
Recoverable Queue
• Separate implementation takes advantage
of semantics to improve performance
– enqueue and dequeue are atomic and isolated,
but some queue locks are released
immediately
acquire lock on queue and entry I1
release lock on queue
release lock on I1
T1: enq(I1) …………..compute………………commit
T2:
enq(I2) …………….compute…………..
T3:
enq(I3) ………………..compute …………
acquire lock
T4: on queue and I
enq(I4)…………….compute ….. 72
2
Recoverable Queue
begin transaction
select
update
enqueue
select
dequeue
commit
DBMS
recoverable
queue
• Queue and DBMS are two separate systems
– Transaction must be committed at both but
• isolation is implemented at the DBMS and applies to the
schedule of requests made to the DBMS only
73
Scheduling
• As a result, any scheduling policy for
accessing the queue might be enforced
– but a FIFO queue might not behave in a FIFO
manner
T1: enq(I1) …commit
restore I1
T2:
enq(I2) …commit
T3:
deq(I1) …
abort
T4:
deq(I2) …commit
74
Performing Real-World Actions
• Problem: A real-world action performed from within a
transaction, T, cannot be rolled back if crash occurs before
commit.
T: begin_transaction;
compute;
update database;
activate device;
commit;
crash
• On recovery after a crash, how can we tell if the action has
occurred?
– ATM example: We do not want to dispense cash twice.
75
Performing Real-World Actions
• Solution: (part 1) T enqueues entry. If T aborts, item
is dequeued; if T commits action executed later
T
queue
T: begin_transaction;
compute;
update database;
enqueue entry;
commit;
TD
device
TD: begin_transaction;
dequeue entry;
activate device;
commit;
• Server executes TD in a loop
– but problem still exists within TD
76
Performing Real-World Actions
T
queue
TD
counter
device
• Solution: (part 2)
– Device maintains read-only counter (hardware) that
is automatically incremented with each action
• Action and increment are assumed to occur atomically
– Server performs:
TD: begin_transaction;
dequeue;
activate device;
record counter in db;
commit;
77
Performing Real-World Actions
• On recovery:
Restore queue and database (value read from
counter) to last commit;
if (device value > recorded value)
then discard head entry;
Restart server;
78
Example of Real World Action
• Suppose the hardware counter and the database counter
were both at 100 before the transaction started
– When the hardware performs its action, it increments its counter to
101
– TD would then increment the database counter to 101
• If the system crashed after the hardware performed its
action the database increment (if it had occurred) would be
rolled back to 100
• Thus when the system recovered
– If the hardware counter was 101 and the database counter was 100,
we would know that the action had been performed.
– If both counters were the same (100), we would know that the
action had not taken place.
79
Forwarding Agent
• Implementing deferred service.
enqueue
client
dequeue
invoke
agent
server
reply
Request
queue
dequeue
Response
queue
enqueue
• In general there are multiple clients (producers) and
multiple servers (consumers)
80
Workflows
• Problem: None of the previous models are
sufficiently flexible to describe complex,
long-running enterprise processes involving
computational and non-computational tasks
in distributed, heterogeneous systems over
extended periods of time
81
Workflow Task
• Self-contained job performed by an agent
– Inventory transaction (agent = database server)
– Packing task (agent = human)
• Has an associated role that defines type of job
– An agent can perform specified roles
• Accepts input from other tasks, produces output
• Has physical status: committed, aborted, ...
– Committed task has logical status: success, failure
82
Workflow
• Task execution precedence specified
separately from task itself
– using control flow language:
initiate T2, T3 when T1 committed
– or using graphical tool:
T2
T1
concurrency
T3
AND condition
83
Workflow
• Conditional alternatives can be specified:
if (condition) execute T1 else execute T2
• Conditions:
– Logical/physical status of a task
– Time of day
– Value of a variable output by a task
• Alternative paths can be specified in case of
task failure
84
Workflow
• Specifies flow of data between tasks
T2
T1
T3
85
Execution Precedence in a
Catalog Ordering System
O
R
bill
OR
AND
take
order
update
remove
package
by
air
shipping
complete
by land
86
Flow of Data in a Catalog
Ordering System
bill
by
air
take
order
update
remove
package
shipping
complete
by land
87
Workflow Agent
• Capable of performing tasks
• Has a set of associated roles describing tasks it can
do
• Has a worklist listing tasks that have been
assigned to it
• Possible implementation:
– Worklist stored in a recoverable queue
– Agent is an infinitely looping process that processes
one queue element on each iteration
88
Workflow and ACID Properties
• Individual tasks might be ACID, but
workflow as a whole is not
– Some task might not be essential: its failure is
ignored even though workflow completes
– Concurrent workflows might see each other’s
intermediate state
– Might not choose to compensate for a task even
though workflow fails
89
Workflow and ACID Properties
• Each task is either
– Retriable: Can ultimately be made to commit if
retried a sufficient number of times (e.g., deposit)
– Compensatable: Compensating task exists (e.g.,
withdraw)
– Pivot: Neither retriable nor compensatable (e.g.,
buy a non-refundable ticket)
90
Workflow and ACID Properties
• The atomicity of a workflow is guaranteed
if each execution path is characterized by
{compensatable}*, [pivot], {retriable}*
• This does not guarantee isolation since
intermediate states are visible
91
Workflow Management System
• Provides mechanism for specifying workflow (control
flow language, GUI)
• Provides mechanism for controlling execution of
concurrent workflows:
–
–
–
–
–
–
–
Roles and agents
Worklists and load balancing
Filters (data reformatting) and controls flow of data
Task activation
Maintain workflow state durably (data, task status)
Use of recoverable queues
Failure recovery of WFMS itself (resume workflows)
92
Importance of Workflows
• Allows management of an enterprise to
guarantee that certain activities are carried
out in accordance with established business
rules, even though those activities involve a
collection of agents, perhaps in different
locations and perhaps with minimal training
93
Scientific Workflows
• Today, many scientific discoveries are achieved
through complex and distributed scientific
computations that are represented and structured
as scientific workflows.
• While business workflows tend to be control and
event driven, scientific workflows are frequently
(not always) data-driven, featuring dataflows.
• Key questions: are business workflow
technologies sufficient for handling scientific
needs?
94
Scientific Workflows Versus
Scientific Workflows
• What is a business workflow and what is a scientific
workflow?
• The boundary is fuzzy, particularly from a technology
point of view.
• Today’s workflow technologies are still far from being
mature for business applications and scientific
applications.
• We do not have to focus on drawing the line between the
two worlds.
• A reasonable vision would be: they have something in
common, but they all differ due to different needs.
95
Key requirements for scientific workflows
• User-oriented workflow management. Scientists should be able to
design, modify, run, re-run, and monitor complex workflows easily
without the burden of managing computation themselves.
• Human cycle optimization. Optimize human cycles in addition to
CPU cycles (how long does it take a scientist to arrive at the right
workflow with the best parameter values and input data?)
• Reproducibility. Enabling technologies are needed to make all
scientific results reproducible.
• Heterogeneous services integration. It should be easy to integrate
services (current and future) from various distributed heterogeneous
environments.
• Data product management. Scientists should be relieved from the
burden of managing large volume of data products and their frequent
movement.
96